Giga and Mapbox join forces with mapping gaming to locate schools across the globe
Information from Giga
When UNICEF and ITU launched Giga in 2019, they set out to connect every school to the internet to ensure children globally had equal access to information, opportunity, and choice – an urgency amplified by the emergence of Covid-19.
The project focuses on three aspects related to connectivity – identify the location and connectivity status of schools, procure broadband connectivity, and provide financial services to fund this activity.
Finding ways to maps schools at scale
The team at Giga soon realised most countries have no single source of school locations. Finding schools is important to establish an anchor for broadband infrastructure to serve communities. Without a single source of school locations, they turned their attention to using satellite imagery to identify potential schools for ground teams to verify. Using artificial intelligence to find schools rendered good initial results, as evidenced by their work in Colombia identifying 6900 previously unmapped schools.
To accelerate their school mapping efforts, Giga collaborated with Mapbox to create The Mapping Game, a crowdsourced application to identify schools using satellite imagery. This app lets anyone look at images to help create a training dataset, by judging possible school locations with simple Yes/No/Unsure answers. The collective assessment of the images helps to refine the machine learning algorithms that the project’s other partners, Development Seed, are developing.
This project accelerated last fall when Mapbox updated their entire global basemap with 50 cm satellite imagery from Maxar. This update enabled the game to work and the project to launch, and to scale the project across the planet.
Mapbox Community was invited to test the game. For their challenge, the game was loaded with 1170 coordinates located across ten countries and four continents – with the Giga team already knowing the true answer (i.e., confirmed a school, not a school, or unclear) based on work done with Development Seed to develop initial machine learning models, and the manual validation efforts it required. A smaller number of coordinates were crowdsourced on-the-ground in Niger and Zimbabwe in the past year and included as well.
- Consistent with an initial pilot in 2020, there were more false negatives than false positives, indicating that participants err on the side of caution and/or mistake schools for non-schools.
- Interestingly, there are two outliers to this trend – Kazakhstan and Uzbekistan – where there were far fewer false negatives. In these countries, the satellite imagery is significantly better in terms of spatial resolution than before, and school patterns also tend to be more easily recognisable.
- By contrast, there were significantly more “No” tags for coordinates in Niger than any other country, where satellite imagery was worse.
These results support the hypothesis that the ability to correctly identify a school is dependent on high quality satellite imagery.
Based on more than 40 000 responses received through the mapping challenge, the Giga team used the XGBoost Algorithm to train a machine learning model that tested with 99% accuracy and 0 false negatives.
Opening the mapping game to all
The Mapping Game is now ready to be played by anyone who wishes to contribute to the project. As the project expands the game will be updated with new locations to develop machine learning models that accurately identify schools across all contexts.
Help locate schools with a few clicks at https://game.projectconnect.world